OpenAI unveils first chip as part of Broadcom deal in effort to 'build the full stack'
CNBC ● Covered by 3 sources
OpenAI and Broadcom showed off their first custom AI chip, Jalapeño, built to run ChatGPT more efficiently. It's OpenAI's boldest move yet to stop depending entirely on Nvidia.
Based on reporting by CNBC — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
OpenAI just handed the world a name straight out of a taco truck menu. Jalapeño is the first chip born from its 18-month partnership with Broadcom, and it's not a concept or a slide in a slideshow — a physical sample landed in OpenAI's hands on Wednesday. The chip is built for inference, the unglamorous but massively expensive job of actually running ChatGPT and OpenAI's other products for hundreds of millions of users, rather than training new models from scratch.
What's notable here isn't just the chip itself but the speed. Greg Brockman told CNBC's David Faber that OpenAI designed Jalapeño end to end in nine months, and that its own AI models sped up the process in ways that genuinely caught the team off guard. That's a small, telling detail: an AI company using its own AI to help design the silicon that will run more AI. It's the kind of self-reinforcing loop that sounds almost too neat, except it's happening in public, on CNBC, with dollar figures attached.
The bigger context is OpenAI's insatiable appetite for compute. Brockman didn't mince words, saying the company simply cannot get enough of it. Broadcom CEO Hock Tan backed that up, describing demand from his AI customers — OpenAI among them — as something his company can't fully address even looking three years out, into 2028. That's not typical vendor spin; Broadcom's stock has climbed nearly sevenfold since 2022, so Tan has receipts.
This chip is also OpenAI hedging its Nvidia bet. The company built its entire empire on Nvidia GPUs, but it's now stacking up alternatives: Amazon's Trainium chips, AMD silicon, a deal with Cerebras, and now this custom ASIC with Broadcom. ASICs trade flexibility for cost and efficiency — they're built to do one job very well rather than everything reasonably well, which suits inference at OpenAI's scale. The two companies are framing this as building toward 10 gigawatts of custom compute, which is an almost absurd amount of power to be planning around for chips that haven't shipped yet.
Don't expect Jalapeño to show up in your ChatGPT session anytime soon. Tan described 2026 as small prototype territory, with real scaling starting in 2027 and full production hitting by the first half of 2028. So this is less a product launch than a marker — proof that OpenAI is serious about owning more of its own stack, chip by chip, even if the payoff is still a couple of years out.
My take — AI-written commentary, not fact-checked reporting
Every major AI lab eventually discovers it hates paying rent to Nvidia, and OpenAI just took the most expensive route out: building its own silicon from scratch. I don't think this makes Jalapeño a Nvidia killer — it's an insurance policy, and a smart one, given how badly OpenAI needs compute it can't always buy. But naming your chip after a pepper while quietly planning for 10 gigawatts of power draws a pretty on-brand contrast between Silicon Valley's whimsy and its genuinely staggering infrastructure ambitions.
Read more about this at: CNBC